How to evaluate an SVG dataset before you license it

Run the checks yourself

Every claim a vendor makes about a vector dataset can be verified with an XML parser and a renderer. Ask for a sample large enough to be representative, then measure rather than read the brochure.

Six measurements

In the order that will save you the most time:

• Validity. Parse every file and render every file. Anything that fails to render is not data, and a vendor who has not run this on the whole line cannot tell you the rate.
• Text editability. Count text nodes per file and check that visible copy is among them. Files whose type has been outlined will render perfectly and teach nothing about text.
• Raster content. Count image nodes. A file embedding a bitmap is a mixed signal; decide deliberately whether you want them, and be able to filter on the count.
• Duplicates. Hash the canonicalised document for exact duplicates and hash the render for near-duplicates. Publish your threshold when you report the rate, because the rate is meaningless without it.
• Metadata completeness. Take the vendor's schema and measure how many fields are actually populated across the sample.
• Provenance. Ask where each file came from, in writing, per asset. If the answer is a category rather than a record, that is your answer.

Two questions that reveal most

Ask whether a stated percentage was measured over the whole line or over a sample, and how large the sample was. Ask what the vendor excludes from delivery and why. Both answers tell you how carefully the corpus has actually been examined.